It's essentially impossible to prove in court that training data was obtained or used improperly unless you go and tell on yourself. And even then it requires you to actually make someone with a lot of money mad, or to not have enough money yourself. Certainly microsoft would have already caught lots of flak for training their models on every github repo, instead they got a minor paddling from the public eye that went away after not much time had passed.

You can legally scrape anything that does not require a login in the US. You can also legally train an AI on it for now.

The point is that the data is easily accessible. If you wanted to get your hands on the data while simultaneously keeping them clean, contract with a Russian contracting company to give you a data dump. You don't need to know how they got it.

you're aware openai trained on a boatload of pirated ebooks?

they "steal" access to data because the LLM launders it on the other end